AI-ordlista
Den kompletta ordlistan över AI
t-SNE
Non-linear dimensionality reduction algorithm using probability distributions to preserve local structures when visualizing high-dimensional data.
Perplexity
Crucial t-SNE parameter controlling the effective number of neighbors considered for each point, influencing the balance between local and global structure.
Kullback-Leibler divergence
Cost function used in t-SNE measuring the dissimilarity between probability distributions in high and low-dimensional space.
Similarity matrix
Mathematical structure representing probabilistic relationships between pairs of points in the original space, based on Gaussian distances.
Gaussian kernel
Exponential kernel function used to convert Euclidean distances into conditional probabilities in high-dimensional space.
t-distribution
Heavy-tailed probability distribution used in low-dimensional space to effectively separate similar points from dissimilar ones.
Crowding Problem
Phenomenon where high-dimensional points become compressed in reduced space, solved by t-SNE through the t-distribution.
Barnes-Hut t-SNE
Optimized variant of t-SNE using a quadtree approximation to reduce computational complexity from O(n²) to O(n log n).
Early Exaggeration
Initial phase of t-SNE that artificially amplifies similarities to form well-separated clusters before final refinement.
Gradient Descent
Iterative optimization algorithm that minimizes KL divergence by gradually adjusting positions in the low-dimensional space.
Learning Rate
Parameter controlling the magnitude of position updates at each iteration, crucial for convergence and final quality.
Momentum
Convergence acceleration technique that adds a fraction of the previous gradient to the current gradient in t-SNE optimization.
Local Structure
Preservation of immediate neighborhood relationships between points, a main characteristic of t-SNE unlike global structures.
Conditional Probabilities
Symmetrized similarity measures between points calculated as probabilities that one point chooses another as a neighbor.
Multi-Scale t-SNE
Extension of t-SNE that combines multiple perplexities to simultaneously capture local and global data structures.
Parametric t-SNE
Variant that learns a parametric mapping function allowing projection of new data without full recalculation.
Trustworthiness
Evaluation metric quantifying the preservation of close neighbors in the projection compared to the original space.
Neighborhood Graph
Graph structure representing neighborhood relationships used to initialize and visualize similarities in t-SNE.